From the little that you say, I infer that this is a toy project.
If your segmentation works reliably, you should have large, stable regions. Make sure that you keep only such regions and discard clutter.
A tumor should be distinguished as a region of an unexpected gray level, differing in shape from a normal region.
Assuming that you have a reference image of a normal brain, and also assuming that no registration is required (all images being taken with the heads in exactly the same position), you can compare the regions two by two, for instance using the Jaccard similarity measure (
http://en.wikipedia.org/wiki/Jaccard_coefficient[
^]). It is an easy matter to count all pixels in the interesections and unions (though the pairwise comparison process will be very costly).
Abnormal regions will appear as having a low Jaccard value.